Re: [PATCH] NFSv4.2: fix setattr caching of TIME_[MODIFY|ACCESS]_SET when timestamps are delegated

[Jeff Layton <jlayton@kernel.org>]

On Tue, 2025-05-06 at 17:25 +0300, Sagi Grimberg wrote:
> > > > >    
> > > > > 
> > > > > 
> > > > > 
> > > > FWIW, we've been chasing some problems with the git regression
> > > > testsuite when attribute delegation is enabled. It would be interesting
> > > > to test this patch to see if it changes that behavior.
> > > Can you elaborate? didn't notice that git uses the times ATTR_*_SET
> > > variant too often.
> > > 
> > Unfortunately, not much.
> > 
> > If you turn on attribute delegation, and then run the git regression
> > suite in a highly-threaded configuration, some of the tests fail. I've
> > made a couple of stabs at trying to narrow down a the reproducer, but
> > no luck so far.
> > 
> > My guess is that it's a client-side bug:
> > 
> > The server is fairly simple here -- if there is an outstanding
> > delegation, it asks the delegation holder for attributes via CB_GETATTR
> > and then passes those along to the client.
> > 
> > The client however has traditionally relied on the server to provide
> > updated attributes, rather than handling timestamps itself, and it
> > wouldn't surprise me if it just didn't get all of those places right.
> 
> It's running, not failing yet... Which test cases fail?

I typically run this under kdevops, using the "stress" profile (which
runs 2 threads for every CPU in the system):

jlayton@walgis:~/git/kdevops/workflows/gitr/results/6.15.0-rc5-ga182e4cc7f89/nfs-v42$ cat 20250505T100749.summary

Test Summary Report
-------------------
./t4137-apply-submodule.sh                         (Wstat: 256 (exited 1) Tests: 28 Failed: 1)
  Failed test:  22
  Non-zero exit status: 1
./t5701-git-serve.sh                               (Wstat: 256 (exited 1) Tests: 24 Failed: 2)
  Failed tests:  1, 21
  Non-zero exit status: 1
./t7112-reset-submodule.sh                         (Wstat: 256 (exited 1) Tests: 82 Failed: 3)
  Failed tests:  6-7, 20
  Non-zero exit status: 1
./t1092-sparse-checkout-compatibility.sh           (Wstat: 256 (exited 1) Tests: 98 Failed: 20)
  Failed tests:  9, 12, 14, 21, 27, 31-39, 48, 50-54
  Non-zero exit status: 1
Files=1007, Tests=30920, 12452 wallclock secs (13.38 usr 12.64 sys + 1115.57 cusr 16188.83 csys = 17330.42 CPU)
Result: FAIL

My efforts at narrowing down the testsuite have mostly resulted in
things not failing. It seems to be some combination of the tests, and
the fact that they are all running in parallel.

-- 
Jeff Layton <jlayton@kernel.org>